Scene Modeling of Autonomous Vehicles Avoiding Stationary and Moving Vehicles on Narrow Roads

TL;DR

The paper tackles autonomous navigation on narrow roads where two vehicles cannot meet simultaneously by introducing SM-NR, a framework that identifies meeting gaps through a road width occupancy principle and initializes trajectories via homology classes. It advances trajectory optimization under tight kinematic and collision-avoidance constraints, and evaluates gaps and paths with a hierarchical decision strategy. Key contributions include a detour-based scene model for stationary vehicles, a gap-identification scheme that yields multiple feasible meeting points, and a robust, real-world-validated decision pipeline that demonstrates improved safety and efficiency in diverse scenarios. The approach shows strong potential for reliable operation on constrained roads and informs future work on curved-road handling and risk-aware planning.

Abstract

Navigating narrow roads with oncoming vehicles is a significant challenge that has garnered considerable public interest. These scenarios often involve sections that cannot accommodate two moving vehicles simultaneously due to the presence of stationary vehicles or limited road width. Autonomous vehicles must therefore profoundly comprehend their surroundings to identify passable areas and execute sophisticated maneuvers. To address this issue, this paper presents a comprehensive model for such an intricate scenario. The primary contribution is the principle of road width occupancy minimization, which models the narrow road problem and identifies candidate meeting gaps. Additionally, the concept of homology classes is introduced to help initialize and optimize candidate trajectories, while evaluation strategies are developed to select the optimal gap and most efficient trajectory. Qualitative and quantitative simulations demonstrate that the proposed approach, SM-NR, achieves high scene pass rates, efficient movement, and robust decisions. Experiments conducted in tiny gap scenarios and conflict scenarios reveal that the autonomous vehicle can robustly select meeting gaps and trajectories, compromising flexibly for safety while advancing bravely for efficiency.

Figures (14)

• Figure 1: Illustration of vehicles passing through the narrow road. The autonomous vehicle is expected to proactively pull over and wait at the red gap to avoid moving vehicles approaching from the oncoming direction.
• Figure 2: Framework of our proposed method, SM-NR. Taking road information and vehicle distribution as input, boundaries are outlined, candidate meeting gaps are identified, and the optimal gap is selected. Candidate trajectories are then initialized and optimized, with the optimal one selected and executed.
• Figure 3: Illustration of scene modeling for detouring a stationary vehicle. The autonomous vehicle is assumed to move closely alongside the road boundary and stationary vehicles. (a) A stationary vehicle is simplified as four corner points $(p_{xi}, p_{yi})|_{i=\{1,2,3,4\}}$. To detour around a point $(p_{x1}, p_{y1})$, the autonomous vehicle starts to turn left at $(x_1,y_1)$ to leave the road boundary and then turns right at $(x_2,y_2)$ until it is in parallel with the stationary vehicle at $(x_3,y_3)$. (b) A whole detour around $(p_{x1}, p_{y1})$ is illustrated in red, consisting of three parts: $minus$, $middle$, and $plus$. Another detour around $(p_{x2},p_{y2})$ in blue intersects with the red one, and they are merged smoothly with a green connection with an orientation $\gamma$, to form a complete detour around the entire stationary vehicle. (c) When two stationary vehicles are close, their detours are smoothly connected with a green curve, which is part of a circle centered at $(o_{x4},o_{y4})$.
• Figure 4: Illustration of idenfitying candidate meeting gaps. (a) An expanded boundary of the rear center of the autonomous vehicle, $L_{av}^{rear}$, is obtained by connecting detours around stationary vehicles$L_{av}^{rear}=\{p_{1,1} \to p_{1,2}\} \cup \{p_{2,1} \to p_{2,2}\} \cup ... \{ p_{5,1} \to p_{5,2}\}$. Expanding $L_{av}^{rear}$ according to the vehicle model, another expanded boundary with the autonomous vehicle's model $L_{av}^{model}$ is obtained. Apply the same process to the moving vehicle, $L_{mv}^{rear}$ and $L_{mv}^{model}$ are obtained. Non-meeting areas are identified where a point on $L_{av}^{model}$ is higher than that on $L_{mv}^{model}$. (b) The model of the autonomous vehicle. (c$_1$-c$_2$) When the autonomous vehicle meets the moving vehicle near the non-meeting area $\textbf{p}_1 \Rightarrow \textbf{p}_2$, the autonomous vehicle must wait before $\textbf{p}_1$ until the moving vehicle advances out of the non-meeting area, or the moving vehicle must wait before $\textbf{p}_2$ until the autonomous vehicle advances out of it. (d$_1$-d$_2$) Two additional maneuvers of cutting in and backing up introduce an additional meeting gap $g_{78}$, enabling a more efficient meeting.
• Figure 5: Illustration of the simplified model on the narrow road scenario.